blob: ff2ebe9458a37ef93c7e4908e1b6156e658c4c75 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
16#include <linux/slab.h>
17#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
22#include <linux/file.h>
23#include <linux/writeback.h>
24#include <linux/blkdev.h>
25#include <linux/buffer_head.h> /* for try_to_release_page(),
26 buffer_heads_over_limit */
27#include <linux/mm_inline.h>
28#include <linux/pagevec.h>
29#include <linux/backing-dev.h>
30#include <linux/rmap.h>
31#include <linux/topology.h>
32#include <linux/cpu.h>
33#include <linux/cpuset.h>
34#include <linux/notifier.h>
35#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080036#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070037#include <linux/kthread.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038
39#include <asm/tlbflush.h>
40#include <asm/div64.h>
41
42#include <linux/swapops.h>
43
Nick Piggin0f8053a2006-03-22 00:08:33 -080044#include "internal.h"
45
Linus Torvalds1da177e2005-04-16 15:20:36 -070046struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070047 /* Incremented by the number of inactive pages that were scanned */
48 unsigned long nr_scanned;
49
Linus Torvalds1da177e2005-04-16 15:20:36 -070050 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040051 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070052
53 int may_writepage;
54
Christoph Lameterf1fd1062006-01-18 17:42:30 -080055 /* Can pages be swapped as part of reclaim? */
56 int may_swap;
57
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
59 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
60 * In this context, it doesn't matter that we scan the
61 * whole list at once. */
62 int swap_cluster_max;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070063
64 int swappiness;
Linus Torvalds1da177e2005-04-16 15:20:36 -070065};
66
67/*
68 * The list of shrinker callbacks used by to apply pressure to
69 * ageable caches.
70 */
71struct shrinker {
72 shrinker_t shrinker;
73 struct list_head list;
74 int seeks; /* seeks to recreate an obj */
75 long nr; /* objs pending delete */
76};
77
78#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
79
80#ifdef ARCH_HAS_PREFETCH
81#define prefetch_prev_lru_page(_page, _base, _field) \
82 do { \
83 if ((_page)->lru.prev != _base) { \
84 struct page *prev; \
85 \
86 prev = lru_to_page(&(_page->lru)); \
87 prefetch(&prev->_field); \
88 } \
89 } while (0)
90#else
91#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
92#endif
93
94#ifdef ARCH_HAS_PREFETCHW
95#define prefetchw_prev_lru_page(_page, _base, _field) \
96 do { \
97 if ((_page)->lru.prev != _base) { \
98 struct page *prev; \
99 \
100 prev = lru_to_page(&(_page->lru)); \
101 prefetchw(&prev->_field); \
102 } \
103 } while (0)
104#else
105#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
106#endif
107
108/*
109 * From 0 .. 100. Higher means more swappy.
110 */
111int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700112long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113
114static LIST_HEAD(shrinker_list);
115static DECLARE_RWSEM(shrinker_rwsem);
116
117/*
118 * Add a shrinker callback to be called from the vm
119 */
120struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
121{
122 struct shrinker *shrinker;
123
124 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
125 if (shrinker) {
126 shrinker->shrinker = theshrinker;
127 shrinker->seeks = seeks;
128 shrinker->nr = 0;
129 down_write(&shrinker_rwsem);
130 list_add_tail(&shrinker->list, &shrinker_list);
131 up_write(&shrinker_rwsem);
132 }
133 return shrinker;
134}
135EXPORT_SYMBOL(set_shrinker);
136
137/*
138 * Remove one
139 */
140void remove_shrinker(struct shrinker *shrinker)
141{
142 down_write(&shrinker_rwsem);
143 list_del(&shrinker->list);
144 up_write(&shrinker_rwsem);
145 kfree(shrinker);
146}
147EXPORT_SYMBOL(remove_shrinker);
148
149#define SHRINK_BATCH 128
150/*
151 * Call the shrink functions to age shrinkable caches
152 *
153 * Here we assume it costs one seek to replace a lru page and that it also
154 * takes a seek to recreate a cache object. With this in mind we age equal
155 * percentages of the lru and ageable caches. This should balance the seeks
156 * generated by these structures.
157 *
158 * If the vm encounted mapped pages on the LRU it increase the pressure on
159 * slab to avoid swapping.
160 *
161 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
162 *
163 * `lru_pages' represents the number of on-LRU pages in all the zones which
164 * are eligible for the caller's allocation attempt. It is used for balancing
165 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700166 *
167 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 */
Andrew Morton69e05942006-03-22 00:08:19 -0800169unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
170 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171{
172 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800173 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174
175 if (scanned == 0)
176 scanned = SWAP_CLUSTER_MAX;
177
178 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700179 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180
181 list_for_each_entry(shrinker, &shrinker_list, list) {
182 unsigned long long delta;
183 unsigned long total_scan;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800184 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185
186 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800187 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 do_div(delta, lru_pages + 1);
189 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800190 if (shrinker->nr < 0) {
191 printk(KERN_ERR "%s: nr=%ld\n",
192 __FUNCTION__, shrinker->nr);
193 shrinker->nr = max_pass;
194 }
195
196 /*
197 * Avoid risking looping forever due to too large nr value:
198 * never try to free more than twice the estimate number of
199 * freeable entries.
200 */
201 if (shrinker->nr > max_pass * 2)
202 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
204 total_scan = shrinker->nr;
205 shrinker->nr = 0;
206
207 while (total_scan >= SHRINK_BATCH) {
208 long this_scan = SHRINK_BATCH;
209 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700210 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700212 nr_before = (*shrinker->shrinker)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
214 if (shrink_ret == -1)
215 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700216 if (shrink_ret < nr_before)
217 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700218 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 total_scan -= this_scan;
220
221 cond_resched();
222 }
223
224 shrinker->nr += total_scan;
225 }
226 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700227 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228}
229
230/* Called without lock on whether page is mapped, so answer is unstable */
231static inline int page_mapping_inuse(struct page *page)
232{
233 struct address_space *mapping;
234
235 /* Page is in somebody's page tables. */
236 if (page_mapped(page))
237 return 1;
238
239 /* Be more reluctant to reclaim swapcache than pagecache */
240 if (PageSwapCache(page))
241 return 1;
242
243 mapping = page_mapping(page);
244 if (!mapping)
245 return 0;
246
247 /* File is mmap'd by somebody? */
248 return mapping_mapped(mapping);
249}
250
251static inline int is_page_cache_freeable(struct page *page)
252{
253 return page_count(page) - !!PagePrivate(page) == 2;
254}
255
256static int may_write_to_queue(struct backing_dev_info *bdi)
257{
Christoph Lameter930d9152006-01-08 01:00:47 -0800258 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 return 1;
260 if (!bdi_write_congested(bdi))
261 return 1;
262 if (bdi == current->backing_dev_info)
263 return 1;
264 return 0;
265}
266
267/*
268 * We detected a synchronous write error writing a page out. Probably
269 * -ENOSPC. We need to propagate that into the address_space for a subsequent
270 * fsync(), msync() or close().
271 *
272 * The tricky part is that after writepage we cannot touch the mapping: nothing
273 * prevents it from being freed up. But we have a ref on the page and once
274 * that page is locked, the mapping is pinned.
275 *
276 * We're allowed to run sleeping lock_page() here because we know the caller has
277 * __GFP_FS.
278 */
279static void handle_write_error(struct address_space *mapping,
280 struct page *page, int error)
281{
282 lock_page(page);
283 if (page_mapping(page) == mapping) {
284 if (error == -ENOSPC)
285 set_bit(AS_ENOSPC, &mapping->flags);
286 else
287 set_bit(AS_EIO, &mapping->flags);
288 }
289 unlock_page(page);
290}
291
Christoph Lameter04e62a22006-06-23 02:03:38 -0700292/* possible outcome of pageout() */
293typedef enum {
294 /* failed to write page out, page is locked */
295 PAGE_KEEP,
296 /* move page to the active list, page is locked */
297 PAGE_ACTIVATE,
298 /* page has been sent to the disk successfully, page is unlocked */
299 PAGE_SUCCESS,
300 /* page is clean and locked */
301 PAGE_CLEAN,
302} pageout_t;
303
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304/*
Andrew Morton1742f192006-03-22 00:08:21 -0800305 * pageout is called by shrink_page_list() for each dirty page.
306 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 */
Christoph Lameter04e62a22006-06-23 02:03:38 -0700308static pageout_t pageout(struct page *page, struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309{
310 /*
311 * If the page is dirty, only perform writeback if that write
312 * will be non-blocking. To prevent this allocation from being
313 * stalled by pagecache activity. But note that there may be
314 * stalls if we need to run get_block(). We could test
315 * PagePrivate for that.
316 *
317 * If this process is currently in generic_file_write() against
318 * this page's queue, we can perform writeback even if that
319 * will block.
320 *
321 * If the page is swapcache, write it back even if that would
322 * block, for some throttling. This happens by accident, because
323 * swap_backing_dev_info is bust: it doesn't reflect the
324 * congestion state of the swapdevs. Easy to fix, if needed.
325 * See swapfile.c:page_queue_congested().
326 */
327 if (!is_page_cache_freeable(page))
328 return PAGE_KEEP;
329 if (!mapping) {
330 /*
331 * Some data journaling orphaned pages can have
332 * page->mapping == NULL while being dirty with clean buffers.
333 */
akpm@osdl.org323aca62005-04-16 15:24:06 -0700334 if (PagePrivate(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 if (try_to_free_buffers(page)) {
336 ClearPageDirty(page);
337 printk("%s: orphaned page\n", __FUNCTION__);
338 return PAGE_CLEAN;
339 }
340 }
341 return PAGE_KEEP;
342 }
343 if (mapping->a_ops->writepage == NULL)
344 return PAGE_ACTIVATE;
345 if (!may_write_to_queue(mapping->backing_dev_info))
346 return PAGE_KEEP;
347
348 if (clear_page_dirty_for_io(page)) {
349 int res;
350 struct writeback_control wbc = {
351 .sync_mode = WB_SYNC_NONE,
352 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700353 .range_start = 0,
354 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355 .nonblocking = 1,
356 .for_reclaim = 1,
357 };
358
359 SetPageReclaim(page);
360 res = mapping->a_ops->writepage(page, &wbc);
361 if (res < 0)
362 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800363 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 ClearPageReclaim(page);
365 return PAGE_ACTIVATE;
366 }
367 if (!PageWriteback(page)) {
368 /* synchronous write or broken a_ops? */
369 ClearPageReclaim(page);
370 }
371
372 return PAGE_SUCCESS;
373 }
374
375 return PAGE_CLEAN;
376}
377
Christoph Lameterb20a3502006-03-22 00:09:12 -0800378int remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800379{
380 if (!mapping)
381 return 0; /* truncate got there first */
382
383 write_lock_irq(&mapping->tree_lock);
384
385 /*
386 * The non-racy check for busy page. It is critical to check
387 * PageDirty _after_ making sure that the page is freeable and
388 * not in use by anybody. (pagecache + us == 2)
389 */
390 if (unlikely(page_count(page) != 2))
391 goto cannot_free;
392 smp_rmb();
393 if (unlikely(PageDirty(page)))
394 goto cannot_free;
395
396 if (PageSwapCache(page)) {
397 swp_entry_t swap = { .val = page_private(page) };
398 __delete_from_swap_cache(page);
399 write_unlock_irq(&mapping->tree_lock);
400 swap_free(swap);
401 __put_page(page); /* The pagecache ref */
402 return 1;
403 }
404
405 __remove_from_page_cache(page);
406 write_unlock_irq(&mapping->tree_lock);
407 __put_page(page);
408 return 1;
409
410cannot_free:
411 write_unlock_irq(&mapping->tree_lock);
412 return 0;
413}
414
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415/*
Andrew Morton1742f192006-03-22 00:08:21 -0800416 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 */
Andrew Morton1742f192006-03-22 00:08:21 -0800418static unsigned long shrink_page_list(struct list_head *page_list,
419 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420{
421 LIST_HEAD(ret_pages);
422 struct pagevec freed_pvec;
423 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800424 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425
426 cond_resched();
427
428 pagevec_init(&freed_pvec, 1);
429 while (!list_empty(page_list)) {
430 struct address_space *mapping;
431 struct page *page;
432 int may_enter_fs;
433 int referenced;
434
435 cond_resched();
436
437 page = lru_to_page(page_list);
438 list_del(&page->lru);
439
440 if (TestSetPageLocked(page))
441 goto keep;
442
443 BUG_ON(PageActive(page));
444
445 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800446
447 if (!sc->may_swap && page_mapped(page))
448 goto keep_locked;
449
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 /* Double the slab pressure for mapped and swapcache pages */
451 if (page_mapped(page) || PageSwapCache(page))
452 sc->nr_scanned++;
453
454 if (PageWriteback(page))
455 goto keep_locked;
456
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800457 referenced = page_referenced(page, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 /* In active use or really unfreeable? Activate it. */
459 if (referenced && page_mapping_inuse(page))
460 goto activate_locked;
461
462#ifdef CONFIG_SWAP
463 /*
464 * Anonymous process memory has backing store?
465 * Try to allocate it some swap space here.
466 */
Christoph Lameter6e5ef1a2006-03-22 00:08:45 -0800467 if (PageAnon(page) && !PageSwapCache(page))
Christoph Lameter1480a542006-01-08 01:00:53 -0800468 if (!add_to_swap(page, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470#endif /* CONFIG_SWAP */
471
472 mapping = page_mapping(page);
473 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
474 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
475
476 /*
477 * The page is mapped into the page tables of one or more
478 * processes. Try to unmap it here.
479 */
480 if (page_mapped(page) && mapping) {
Christoph Lametera48d07a2006-02-01 03:05:38 -0800481 switch (try_to_unmap(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 case SWAP_FAIL:
483 goto activate_locked;
484 case SWAP_AGAIN:
485 goto keep_locked;
486 case SWAP_SUCCESS:
487 ; /* try to free the page below */
488 }
489 }
490
491 if (PageDirty(page)) {
492 if (referenced)
493 goto keep_locked;
494 if (!may_enter_fs)
495 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800496 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 goto keep_locked;
498
499 /* Page is dirty, try to write it out here */
500 switch(pageout(page, mapping)) {
501 case PAGE_KEEP:
502 goto keep_locked;
503 case PAGE_ACTIVATE:
504 goto activate_locked;
505 case PAGE_SUCCESS:
506 if (PageWriteback(page) || PageDirty(page))
507 goto keep;
508 /*
509 * A synchronous write - probably a ramdisk. Go
510 * ahead and try to reclaim the page.
511 */
512 if (TestSetPageLocked(page))
513 goto keep;
514 if (PageDirty(page) || PageWriteback(page))
515 goto keep_locked;
516 mapping = page_mapping(page);
517 case PAGE_CLEAN:
518 ; /* try to free the page below */
519 }
520 }
521
522 /*
523 * If the page has buffers, try to free the buffer mappings
524 * associated with this page. If we succeed we try to free
525 * the page as well.
526 *
527 * We do this even if the page is PageDirty().
528 * try_to_release_page() does not perform I/O, but it is
529 * possible for a page to have PageDirty set, but it is actually
530 * clean (all its buffers are clean). This happens if the
531 * buffers were written out directly, with submit_bh(). ext3
532 * will do this, as well as the blockdev mapping.
533 * try_to_release_page() will discover that cleanness and will
534 * drop the buffers and mark the page clean - it can be freed.
535 *
536 * Rarely, pages can have buffers and no ->mapping. These are
537 * the pages which were not successfully invalidated in
538 * truncate_complete_page(). We try to drop those buffers here
539 * and if that worked, and the page is no longer mapped into
540 * process address space (page_count == 1) it can be freed.
541 * Otherwise, leave the page on the LRU so it is swappable.
542 */
543 if (PagePrivate(page)) {
544 if (!try_to_release_page(page, sc->gfp_mask))
545 goto activate_locked;
546 if (!mapping && page_count(page) == 1)
547 goto free_it;
548 }
549
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800550 if (!remove_mapping(mapping, page))
551 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552
553free_it:
554 unlock_page(page);
Andrew Morton05ff5132006-03-22 00:08:20 -0800555 nr_reclaimed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556 if (!pagevec_add(&freed_pvec, page))
557 __pagevec_release_nonlru(&freed_pvec);
558 continue;
559
560activate_locked:
561 SetPageActive(page);
562 pgactivate++;
563keep_locked:
564 unlock_page(page);
565keep:
566 list_add(&page->lru, &ret_pages);
567 BUG_ON(PageLRU(page));
568 }
569 list_splice(&ret_pages, page_list);
570 if (pagevec_count(&freed_pvec))
571 __pagevec_release_nonlru(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700572 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800573 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574}
575
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800576/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 * zone->lru_lock is heavily contended. Some of the functions that
578 * shrink the lists perform better by taking out a batch of pages
579 * and working on them outside the LRU lock.
580 *
581 * For pagecache intensive workloads, this function is the hottest
582 * spot in the kernel (apart from copy_*_user functions).
583 *
584 * Appropriate locks must be held before calling this function.
585 *
586 * @nr_to_scan: The number of pages to look through on the list.
587 * @src: The LRU list to pull pages off.
588 * @dst: The temp list to put pages on to.
589 * @scanned: The number of pages that were scanned.
590 *
591 * returns how many pages were moved onto *@dst.
592 */
Andrew Morton69e05942006-03-22 00:08:19 -0800593static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
594 struct list_head *src, struct list_head *dst,
595 unsigned long *scanned)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596{
Andrew Morton69e05942006-03-22 00:08:19 -0800597 unsigned long nr_taken = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 struct page *page;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800599 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800601 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800602 struct list_head *target;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 page = lru_to_page(src);
604 prefetchw_prev_lru_page(page, src, flags);
605
Nick Piggin8d438f92006-03-22 00:07:59 -0800606 BUG_ON(!PageLRU(page));
607
Nick Piggin053837f2006-01-18 17:42:27 -0800608 list_del(&page->lru);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800609 target = src;
610 if (likely(get_page_unless_zero(page))) {
Nick Piggin053837f2006-01-18 17:42:27 -0800611 /*
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800612 * Be careful not to clear PageLRU until after we're
613 * sure the page is not being freed elsewhere -- the
614 * page release code relies on it.
Nick Piggin053837f2006-01-18 17:42:27 -0800615 */
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800616 ClearPageLRU(page);
617 target = dst;
618 nr_taken++;
619 } /* else it is being freed elsewhere */
Nick Piggin46453a62006-03-22 00:07:58 -0800620
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800621 list_add(&page->lru, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 }
623
624 *scanned = scan;
625 return nr_taken;
626}
627
628/*
Andrew Morton1742f192006-03-22 00:08:21 -0800629 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
630 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631 */
Andrew Morton1742f192006-03-22 00:08:21 -0800632static unsigned long shrink_inactive_list(unsigned long max_scan,
633 struct zone *zone, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634{
635 LIST_HEAD(page_list);
636 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -0800637 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800638 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
640 pagevec_init(&pvec, 1);
641
642 lru_add_drain();
643 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -0800644 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -0800646 unsigned long nr_taken;
647 unsigned long nr_scan;
648 unsigned long nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649
650 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
651 &zone->inactive_list,
652 &page_list, &nr_scan);
653 zone->nr_inactive -= nr_taken;
654 zone->pages_scanned += nr_scan;
655 spin_unlock_irq(&zone->lru_lock);
656
Andrew Morton69e05942006-03-22 00:08:19 -0800657 nr_scanned += nr_scan;
Andrew Morton1742f192006-03-22 00:08:21 -0800658 nr_freed = shrink_page_list(&page_list, sc);
Andrew Morton05ff5132006-03-22 00:08:20 -0800659 nr_reclaimed += nr_freed;
Nick Piggina74609f2006-01-06 00:11:20 -0800660 local_irq_disable();
661 if (current_is_kswapd()) {
Christoph Lameterf8891e52006-06-30 01:55:45 -0700662 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
663 __count_vm_events(KSWAPD_STEAL, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -0800664 } else
Christoph Lameterf8891e52006-06-30 01:55:45 -0700665 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
666 __count_vm_events(PGACTIVATE, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -0800667
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800668 if (nr_taken == 0)
669 goto done;
670
Nick Piggina74609f2006-01-06 00:11:20 -0800671 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 /*
673 * Put back any unfreeable pages.
674 */
675 while (!list_empty(&page_list)) {
676 page = lru_to_page(&page_list);
Nick Piggin8d438f92006-03-22 00:07:59 -0800677 BUG_ON(PageLRU(page));
678 SetPageLRU(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 list_del(&page->lru);
680 if (PageActive(page))
681 add_page_to_active_list(zone, page);
682 else
683 add_page_to_inactive_list(zone, page);
684 if (!pagevec_add(&pvec, page)) {
685 spin_unlock_irq(&zone->lru_lock);
686 __pagevec_release(&pvec);
687 spin_lock_irq(&zone->lru_lock);
688 }
689 }
Andrew Morton69e05942006-03-22 00:08:19 -0800690 } while (nr_scanned < max_scan);
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800691 spin_unlock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692done:
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800693 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -0800695 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696}
697
698/*
699 * This moves pages from the active list to the inactive list.
700 *
701 * We move them the other way if the page is referenced by one or more
702 * processes, from rmap.
703 *
704 * If the pages are mostly unmapped, the processing is fast and it is
705 * appropriate to hold zone->lru_lock across the whole operation. But if
706 * the pages are mapped, the processing is slow (page_referenced()) so we
707 * should drop zone->lru_lock around each page. It's impossible to balance
708 * this, so instead we remove the pages from the LRU while processing them.
709 * It is safe to rely on PG_active against the non-LRU pages in here because
710 * nobody will play with that bit on a non-LRU page.
711 *
712 * The downside is that we have to touch page->_count against each page.
713 * But we had to alter page->flags anyway.
714 */
Andrew Morton1742f192006-03-22 00:08:21 -0800715static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
716 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717{
Andrew Morton69e05942006-03-22 00:08:19 -0800718 unsigned long pgmoved;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 int pgdeactivate = 0;
Andrew Morton69e05942006-03-22 00:08:19 -0800720 unsigned long pgscanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 LIST_HEAD(l_hold); /* The pages which were snipped off */
722 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
723 LIST_HEAD(l_active); /* Pages to go onto the active_list */
724 struct page *page;
725 struct pagevec pvec;
726 int reclaim_mapped = 0;
Christoph Lameter2903fb12006-02-11 17:55:55 -0800727
Christoph Lameter6e5ef1a2006-03-22 00:08:45 -0800728 if (sc->may_swap) {
Christoph Lameter2903fb12006-02-11 17:55:55 -0800729 long mapped_ratio;
730 long distress;
731 long swap_tendency;
732
733 /*
734 * `distress' is a measure of how much trouble we're having
735 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
736 */
737 distress = 100 >> zone->prev_priority;
738
739 /*
740 * The point of this algorithm is to decide when to start
741 * reclaiming mapped memory instead of just pagecache. Work out
742 * how much memory
743 * is mapped.
744 */
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700745 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
746 global_page_state(NR_ANON_PAGES)) * 100) /
Christoph Lameterbf02cf42006-06-30 01:55:36 -0700747 vm_total_pages;
Christoph Lameter2903fb12006-02-11 17:55:55 -0800748
749 /*
750 * Now decide how much we really want to unmap some pages. The
751 * mapped ratio is downgraded - just because there's a lot of
752 * mapped memory doesn't necessarily mean that page reclaim
753 * isn't succeeding.
754 *
755 * The distress ratio is important - we don't want to start
756 * going oom.
757 *
758 * A 100% value of vm_swappiness overrides this algorithm
759 * altogether.
760 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -0700761 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
Christoph Lameter2903fb12006-02-11 17:55:55 -0800762
763 /*
764 * Now use this metric to decide whether to start moving mapped
765 * memory onto the inactive list.
766 */
767 if (swap_tendency >= 100)
768 reclaim_mapped = 1;
769 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770
771 lru_add_drain();
772 spin_lock_irq(&zone->lru_lock);
773 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
774 &l_hold, &pgscanned);
775 zone->pages_scanned += pgscanned;
776 zone->nr_active -= pgmoved;
777 spin_unlock_irq(&zone->lru_lock);
778
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 while (!list_empty(&l_hold)) {
780 cond_resched();
781 page = lru_to_page(&l_hold);
782 list_del(&page->lru);
783 if (page_mapped(page)) {
784 if (!reclaim_mapped ||
785 (total_swap_pages == 0 && PageAnon(page)) ||
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800786 page_referenced(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 list_add(&page->lru, &l_active);
788 continue;
789 }
790 }
791 list_add(&page->lru, &l_inactive);
792 }
793
794 pagevec_init(&pvec, 1);
795 pgmoved = 0;
796 spin_lock_irq(&zone->lru_lock);
797 while (!list_empty(&l_inactive)) {
798 page = lru_to_page(&l_inactive);
799 prefetchw_prev_lru_page(page, &l_inactive, flags);
Nick Piggin8d438f92006-03-22 00:07:59 -0800800 BUG_ON(PageLRU(page));
801 SetPageLRU(page);
Nick Piggin4c84cac2006-03-22 00:08:00 -0800802 BUG_ON(!PageActive(page));
803 ClearPageActive(page);
804
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 list_move(&page->lru, &zone->inactive_list);
806 pgmoved++;
807 if (!pagevec_add(&pvec, page)) {
808 zone->nr_inactive += pgmoved;
809 spin_unlock_irq(&zone->lru_lock);
810 pgdeactivate += pgmoved;
811 pgmoved = 0;
812 if (buffer_heads_over_limit)
813 pagevec_strip(&pvec);
814 __pagevec_release(&pvec);
815 spin_lock_irq(&zone->lru_lock);
816 }
817 }
818 zone->nr_inactive += pgmoved;
819 pgdeactivate += pgmoved;
820 if (buffer_heads_over_limit) {
821 spin_unlock_irq(&zone->lru_lock);
822 pagevec_strip(&pvec);
823 spin_lock_irq(&zone->lru_lock);
824 }
825
826 pgmoved = 0;
827 while (!list_empty(&l_active)) {
828 page = lru_to_page(&l_active);
829 prefetchw_prev_lru_page(page, &l_active, flags);
Nick Piggin8d438f92006-03-22 00:07:59 -0800830 BUG_ON(PageLRU(page));
831 SetPageLRU(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 BUG_ON(!PageActive(page));
833 list_move(&page->lru, &zone->active_list);
834 pgmoved++;
835 if (!pagevec_add(&pvec, page)) {
836 zone->nr_active += pgmoved;
837 pgmoved = 0;
838 spin_unlock_irq(&zone->lru_lock);
839 __pagevec_release(&pvec);
840 spin_lock_irq(&zone->lru_lock);
841 }
842 }
843 zone->nr_active += pgmoved;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844
Christoph Lameterf8891e52006-06-30 01:55:45 -0700845 __count_zone_vm_events(PGREFILL, zone, pgscanned);
846 __count_vm_events(PGDEACTIVATE, pgdeactivate);
847 spin_unlock_irq(&zone->lru_lock);
Nick Piggina74609f2006-01-06 00:11:20 -0800848
849 pagevec_release(&pvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850}
851
852/*
853 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
854 */
Andrew Morton05ff5132006-03-22 00:08:20 -0800855static unsigned long shrink_zone(int priority, struct zone *zone,
856 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857{
858 unsigned long nr_active;
859 unsigned long nr_inactive;
Christoph Lameter86959492006-03-22 00:08:18 -0800860 unsigned long nr_to_scan;
Andrew Morton05ff5132006-03-22 00:08:20 -0800861 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862
Martin Hicks53e9a612005-09-03 15:54:51 -0700863 atomic_inc(&zone->reclaim_in_progress);
864
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865 /*
866 * Add one to `nr_to_scan' just to make sure that the kernel will
867 * slowly sift through the active list.
868 */
Christoph Lameter86959492006-03-22 00:08:18 -0800869 zone->nr_scan_active += (zone->nr_active >> priority) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 nr_active = zone->nr_scan_active;
871 if (nr_active >= sc->swap_cluster_max)
872 zone->nr_scan_active = 0;
873 else
874 nr_active = 0;
875
Christoph Lameter86959492006-03-22 00:08:18 -0800876 zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 nr_inactive = zone->nr_scan_inactive;
878 if (nr_inactive >= sc->swap_cluster_max)
879 zone->nr_scan_inactive = 0;
880 else
881 nr_inactive = 0;
882
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 while (nr_active || nr_inactive) {
884 if (nr_active) {
Christoph Lameter86959492006-03-22 00:08:18 -0800885 nr_to_scan = min(nr_active,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 (unsigned long)sc->swap_cluster_max);
Christoph Lameter86959492006-03-22 00:08:18 -0800887 nr_active -= nr_to_scan;
Andrew Morton1742f192006-03-22 00:08:21 -0800888 shrink_active_list(nr_to_scan, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 }
890
891 if (nr_inactive) {
Christoph Lameter86959492006-03-22 00:08:18 -0800892 nr_to_scan = min(nr_inactive,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 (unsigned long)sc->swap_cluster_max);
Christoph Lameter86959492006-03-22 00:08:18 -0800894 nr_inactive -= nr_to_scan;
Andrew Morton1742f192006-03-22 00:08:21 -0800895 nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
896 sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 }
898 }
899
900 throttle_vm_writeout();
Martin Hicks53e9a612005-09-03 15:54:51 -0700901
902 atomic_dec(&zone->reclaim_in_progress);
Andrew Morton05ff5132006-03-22 00:08:20 -0800903 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904}
905
906/*
907 * This is the direct reclaim path, for page-allocating processes. We only
908 * try to reclaim pages from zones which will satisfy the caller's allocation
909 * request.
910 *
911 * We reclaim from a zone even if that zone is over pages_high. Because:
912 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
913 * allocation or
914 * b) The zones may be over pages_high but they must go *over* pages_high to
915 * satisfy the `incremental min' zone defense algorithm.
916 *
917 * Returns the number of reclaimed pages.
918 *
919 * If a zone is deemed to be full of pinned pages then just give it a light
920 * scan then give up on it.
921 */
Andrew Morton1742f192006-03-22 00:08:21 -0800922static unsigned long shrink_zones(int priority, struct zone **zones,
Andrew Morton05ff5132006-03-22 00:08:20 -0800923 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924{
Andrew Morton05ff5132006-03-22 00:08:20 -0800925 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 int i;
927
928 for (i = 0; zones[i] != NULL; i++) {
929 struct zone *zone = zones[i];
930
Con Kolivasf3fe6512006-01-06 00:11:15 -0800931 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 continue;
933
Paul Jackson9bf22292005-09-06 15:18:12 -0700934 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 continue;
936
Christoph Lameter86959492006-03-22 00:08:18 -0800937 zone->temp_priority = priority;
938 if (zone->prev_priority > priority)
939 zone->prev_priority = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940
Christoph Lameter86959492006-03-22 00:08:18 -0800941 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942 continue; /* Let kswapd poll it */
943
Andrew Morton05ff5132006-03-22 00:08:20 -0800944 nr_reclaimed += shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 }
Andrew Morton05ff5132006-03-22 00:08:20 -0800946 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947}
948
949/*
950 * This is the main entry point to direct page reclaim.
951 *
952 * If a full scan of the inactive list fails to free enough memory then we
953 * are "out of memory" and something needs to be killed.
954 *
955 * If the caller is !__GFP_FS then the probability of a failure is reasonably
956 * high - the zone may be full of dirty or under-writeback pages, which this
957 * caller can't do much about. We kick pdflush and take explicit naps in the
958 * hope that some of these pages can be written. But if the allocating task
959 * holds filesystem locks which prevent writeout this might not work, and the
960 * allocation attempt will fail.
961 */
Andrew Morton69e05942006-03-22 00:08:19 -0800962unsigned long try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963{
964 int priority;
965 int ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -0800966 unsigned long total_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800967 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 unsigned long lru_pages = 0;
970 int i;
Andrew Morton179e9632006-03-22 00:08:18 -0800971 struct scan_control sc = {
972 .gfp_mask = gfp_mask,
973 .may_writepage = !laptop_mode,
974 .swap_cluster_max = SWAP_CLUSTER_MAX,
975 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -0700976 .swappiness = vm_swappiness,
Andrew Morton179e9632006-03-22 00:08:18 -0800977 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978
Christoph Lameterf8891e52006-06-30 01:55:45 -0700979 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980
981 for (i = 0; zones[i] != NULL; i++) {
982 struct zone *zone = zones[i];
983
Paul Jackson9bf22292005-09-06 15:18:12 -0700984 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 continue;
986
987 zone->temp_priority = DEF_PRIORITY;
988 lru_pages += zone->nr_active + zone->nr_inactive;
989 }
990
991 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 sc.nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -0800993 if (!priority)
994 disable_swap_token();
Andrew Morton1742f192006-03-22 00:08:21 -0800995 nr_reclaimed += shrink_zones(priority, zones, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
997 if (reclaim_state) {
Andrew Morton05ff5132006-03-22 00:08:20 -0800998 nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 reclaim_state->reclaimed_slab = 0;
1000 }
1001 total_scanned += sc.nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001002 if (nr_reclaimed >= sc.swap_cluster_max) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003 ret = 1;
1004 goto out;
1005 }
1006
1007 /*
1008 * Try to write back as many pages as we just scanned. This
1009 * tends to cause slow streaming writers to write data to the
1010 * disk smoothly, at the dirtying rate, which is nice. But
1011 * that's undesirable in laptop mode, where we *want* lumpy
1012 * writeout. So in laptop mode, write out the whole world.
1013 */
Andrew Morton179e9632006-03-22 00:08:18 -08001014 if (total_scanned > sc.swap_cluster_max +
1015 sc.swap_cluster_max / 2) {
Pekka J Enberg687a21c2005-06-28 20:44:55 -07001016 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 sc.may_writepage = 1;
1018 }
1019
1020 /* Take a nap, wait for some writeback to complete */
1021 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1022 blk_congestion_wait(WRITE, HZ/10);
1023 }
1024out:
1025 for (i = 0; zones[i] != 0; i++) {
1026 struct zone *zone = zones[i];
1027
Paul Jackson9bf22292005-09-06 15:18:12 -07001028 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 continue;
1030
1031 zone->prev_priority = zone->temp_priority;
1032 }
1033 return ret;
1034}
1035
1036/*
1037 * For kswapd, balance_pgdat() will work across all this node's zones until
1038 * they are all at pages_high.
1039 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 * Returns the number of pages which were actually freed.
1041 *
1042 * There is special handling here for zones which are full of pinned pages.
1043 * This can happen if the pages are all mlocked, or if they are all used by
1044 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1045 * What we do is to detect the case where all pages in the zone have been
1046 * scanned twice and there has been zero successful reclaim. Mark the zone as
1047 * dead and from now on, only perform a short scan. Basically we're polling
1048 * the zone for when the problem goes away.
1049 *
1050 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1051 * zones which have free_pages > pages_high, but once a zone is found to have
1052 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1053 * of the number of free pages in the lower zones. This interoperates with
1054 * the page allocator fallback scheme to ensure that aging of pages is balanced
1055 * across the zones.
1056 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001057static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059 int all_zones_ok;
1060 int priority;
1061 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001062 unsigned long total_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001063 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001065 struct scan_control sc = {
1066 .gfp_mask = GFP_KERNEL,
1067 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001068 .swap_cluster_max = SWAP_CLUSTER_MAX,
1069 .swappiness = vm_swappiness,
Andrew Morton179e9632006-03-22 00:08:18 -08001070 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071
1072loop_again:
1073 total_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001074 nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001075 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001076 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077
1078 for (i = 0; i < pgdat->nr_zones; i++) {
1079 struct zone *zone = pgdat->node_zones + i;
1080
1081 zone->temp_priority = DEF_PRIORITY;
1082 }
1083
1084 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1085 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1086 unsigned long lru_pages = 0;
1087
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001088 /* The swap token gets in the way of swapout... */
1089 if (!priority)
1090 disable_swap_token();
1091
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 all_zones_ok = 1;
1093
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001094 /*
1095 * Scan in the highmem->dma direction for the highest
1096 * zone which needs scanning
1097 */
1098 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1099 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001101 if (!populated_zone(zone))
1102 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001104 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1105 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001107 if (!zone_watermark_ok(zone, order, zone->pages_high,
1108 0, 0)) {
1109 end_zone = i;
1110 goto scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001113 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114scan:
1115 for (i = 0; i <= end_zone; i++) {
1116 struct zone *zone = pgdat->node_zones + i;
1117
1118 lru_pages += zone->nr_active + zone->nr_inactive;
1119 }
1120
1121 /*
1122 * Now scan the zone in the dma->highmem direction, stopping
1123 * at the last zone which needs scanning.
1124 *
1125 * We do this because the page allocator works in the opposite
1126 * direction. This prevents the page allocator from allocating
1127 * pages behind kswapd's direction of progress, which would
1128 * cause too much scanning of the lower zones.
1129 */
1130 for (i = 0; i <= end_zone; i++) {
1131 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001132 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133
Con Kolivasf3fe6512006-01-06 00:11:15 -08001134 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 continue;
1136
1137 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1138 continue;
1139
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001140 if (!zone_watermark_ok(zone, order, zone->pages_high,
1141 end_zone, 0))
1142 all_zones_ok = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 zone->temp_priority = priority;
1144 if (zone->prev_priority > priority)
1145 zone->prev_priority = priority;
1146 sc.nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001147 nr_reclaimed += shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001149 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1150 lru_pages);
Andrew Morton05ff5132006-03-22 00:08:20 -08001151 nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 total_scanned += sc.nr_scanned;
1153 if (zone->all_unreclaimable)
1154 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001155 if (nr_slab == 0 && zone->pages_scanned >=
1156 (zone->nr_active + zone->nr_inactive) * 4)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 zone->all_unreclaimable = 1;
1158 /*
1159 * If we've done a decent amount of scanning and
1160 * the reclaim ratio is low, start doing writepage
1161 * even in laptop mode
1162 */
1163 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Andrew Morton05ff5132006-03-22 00:08:20 -08001164 total_scanned > nr_reclaimed + nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165 sc.may_writepage = 1;
1166 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 if (all_zones_ok)
1168 break; /* kswapd: all done */
1169 /*
1170 * OK, kswapd is getting into trouble. Take a nap, then take
1171 * another pass across the zones.
1172 */
1173 if (total_scanned && priority < DEF_PRIORITY - 2)
1174 blk_congestion_wait(WRITE, HZ/10);
1175
1176 /*
1177 * We do this so kswapd doesn't build up large priorities for
1178 * example when it is freeing in parallel with allocators. It
1179 * matches the direct reclaim path behaviour in terms of impact
1180 * on zone->*_priority.
1181 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001182 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 break;
1184 }
1185out:
1186 for (i = 0; i < pgdat->nr_zones; i++) {
1187 struct zone *zone = pgdat->node_zones + i;
1188
1189 zone->prev_priority = zone->temp_priority;
1190 }
1191 if (!all_zones_ok) {
1192 cond_resched();
1193 goto loop_again;
1194 }
1195
Andrew Morton05ff5132006-03-22 00:08:20 -08001196 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197}
1198
1199/*
1200 * The background pageout daemon, started as a kernel thread
1201 * from the init process.
1202 *
1203 * This basically trickles out pages so that we have _some_
1204 * free memory available even if there is no other activity
1205 * that frees anything up. This is needed for things like routing
1206 * etc, where we otherwise might have all activity going on in
1207 * asynchronous contexts that cannot page things out.
1208 *
1209 * If there are applications that are active memory-allocators
1210 * (most normal use), this basically shouldn't matter.
1211 */
1212static int kswapd(void *p)
1213{
1214 unsigned long order;
1215 pg_data_t *pgdat = (pg_data_t*)p;
1216 struct task_struct *tsk = current;
1217 DEFINE_WAIT(wait);
1218 struct reclaim_state reclaim_state = {
1219 .reclaimed_slab = 0,
1220 };
1221 cpumask_t cpumask;
1222
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 cpumask = node_to_cpumask(pgdat->node_id);
1224 if (!cpus_empty(cpumask))
1225 set_cpus_allowed(tsk, cpumask);
1226 current->reclaim_state = &reclaim_state;
1227
1228 /*
1229 * Tell the memory management that we're a "memory allocator",
1230 * and that if we need more memory we should get access to it
1231 * regardless (see "__alloc_pages()"). "kswapd" should
1232 * never get caught in the normal page freeing logic.
1233 *
1234 * (Kswapd normally doesn't need memory anyway, but sometimes
1235 * you need a small amount of memory in order to be able to
1236 * page out something else, and this flag essentially protects
1237 * us from recursively trying to free more memory as we're
1238 * trying to free the first piece of memory in the first place).
1239 */
Christoph Lameter930d9152006-01-08 01:00:47 -08001240 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241
1242 order = 0;
1243 for ( ; ; ) {
1244 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07001245
1246 try_to_freeze();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247
1248 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1249 new_order = pgdat->kswapd_max_order;
1250 pgdat->kswapd_max_order = 0;
1251 if (order < new_order) {
1252 /*
1253 * Don't sleep if someone wants a larger 'order'
1254 * allocation
1255 */
1256 order = new_order;
1257 } else {
1258 schedule();
1259 order = pgdat->kswapd_max_order;
1260 }
1261 finish_wait(&pgdat->kswapd_wait, &wait);
1262
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001263 balance_pgdat(pgdat, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 }
1265 return 0;
1266}
1267
1268/*
1269 * A zone is low on free memory, so wake its kswapd task to service it.
1270 */
1271void wakeup_kswapd(struct zone *zone, int order)
1272{
1273 pg_data_t *pgdat;
1274
Con Kolivasf3fe6512006-01-06 00:11:15 -08001275 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 return;
1277
1278 pgdat = zone->zone_pgdat;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001279 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 return;
1281 if (pgdat->kswapd_max_order < order)
1282 pgdat->kswapd_max_order = order;
Paul Jackson9bf22292005-09-06 15:18:12 -07001283 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001285 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001287 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288}
1289
1290#ifdef CONFIG_PM
1291/*
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001292 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1293 * from LRU lists system-wide, for given pass and priority, and returns the
1294 * number of reclaimed pages
1295 *
1296 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1297 */
1298static unsigned long shrink_all_zones(unsigned long nr_pages, int pass,
1299 int prio, struct scan_control *sc)
1300{
1301 struct zone *zone;
1302 unsigned long nr_to_scan, ret = 0;
1303
1304 for_each_zone(zone) {
1305
1306 if (!populated_zone(zone))
1307 continue;
1308
1309 if (zone->all_unreclaimable && prio != DEF_PRIORITY)
1310 continue;
1311
1312 /* For pass = 0 we don't shrink the active list */
1313 if (pass > 0) {
1314 zone->nr_scan_active += (zone->nr_active >> prio) + 1;
1315 if (zone->nr_scan_active >= nr_pages || pass > 3) {
1316 zone->nr_scan_active = 0;
1317 nr_to_scan = min(nr_pages, zone->nr_active);
1318 shrink_active_list(nr_to_scan, zone, sc);
1319 }
1320 }
1321
1322 zone->nr_scan_inactive += (zone->nr_inactive >> prio) + 1;
1323 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1324 zone->nr_scan_inactive = 0;
1325 nr_to_scan = min(nr_pages, zone->nr_inactive);
1326 ret += shrink_inactive_list(nr_to_scan, zone, sc);
1327 if (ret >= nr_pages)
1328 return ret;
1329 }
1330 }
1331
1332 return ret;
1333}
1334
1335/*
1336 * Try to free `nr_pages' of memory, system-wide, and return the number of
1337 * freed pages.
1338 *
1339 * Rather than trying to age LRUs the aim is to preserve the overall
1340 * LRU order by reclaiming preferentially
1341 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 */
Andrew Morton69e05942006-03-22 00:08:19 -08001343unsigned long shrink_all_memory(unsigned long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001345 unsigned long lru_pages, nr_slab;
Andrew Morton69e05942006-03-22 00:08:19 -08001346 unsigned long ret = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001347 int pass;
1348 struct reclaim_state reclaim_state;
1349 struct zone *zone;
1350 struct scan_control sc = {
1351 .gfp_mask = GFP_KERNEL,
1352 .may_swap = 0,
1353 .swap_cluster_max = nr_pages,
1354 .may_writepage = 1,
1355 .swappiness = vm_swappiness,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 };
1357
1358 current->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08001359
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001360 lru_pages = 0;
1361 for_each_zone(zone)
1362 lru_pages += zone->nr_active + zone->nr_inactive;
1363
Christoph Lameter9a865ff2006-06-30 01:55:38 -07001364 nr_slab = global_page_state(NR_SLAB);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001365 /* If slab caches are huge, it's better to hit them first */
1366 while (nr_slab >= lru_pages) {
1367 reclaim_state.reclaimed_slab = 0;
1368 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1369 if (!reclaim_state.reclaimed_slab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 break;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001371
1372 ret += reclaim_state.reclaimed_slab;
1373 if (ret >= nr_pages)
1374 goto out;
1375
1376 nr_slab -= reclaim_state.reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001378
1379 /*
1380 * We try to shrink LRUs in 5 passes:
1381 * 0 = Reclaim from inactive_list only
1382 * 1 = Reclaim from active list but don't reclaim mapped
1383 * 2 = 2nd pass of type 1
1384 * 3 = Reclaim mapped (normal reclaim)
1385 * 4 = 2nd pass of type 3
1386 */
1387 for (pass = 0; pass < 5; pass++) {
1388 int prio;
1389
1390 /* Needed for shrinking slab caches later on */
1391 if (!lru_pages)
1392 for_each_zone(zone) {
1393 lru_pages += zone->nr_active;
1394 lru_pages += zone->nr_inactive;
1395 }
1396
1397 /* Force reclaiming mapped pages in the passes #3 and #4 */
1398 if (pass > 2) {
1399 sc.may_swap = 1;
1400 sc.swappiness = 100;
1401 }
1402
1403 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1404 unsigned long nr_to_scan = nr_pages - ret;
1405
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001406 sc.nr_scanned = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001407 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1408 if (ret >= nr_pages)
1409 goto out;
1410
1411 reclaim_state.reclaimed_slab = 0;
1412 shrink_slab(sc.nr_scanned, sc.gfp_mask, lru_pages);
1413 ret += reclaim_state.reclaimed_slab;
1414 if (ret >= nr_pages)
1415 goto out;
1416
1417 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1418 blk_congestion_wait(WRITE, HZ / 10);
1419 }
1420
1421 lru_pages = 0;
Rafael J. Wysocki248a0302006-03-22 00:09:04 -08001422 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001423
1424 /*
1425 * If ret = 0, we could not shrink LRUs, but there may be something
1426 * in slab caches
1427 */
1428 if (!ret)
1429 do {
1430 reclaim_state.reclaimed_slab = 0;
1431 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1432 ret += reclaim_state.reclaimed_slab;
1433 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1434
1435out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 current->reclaim_state = NULL;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001437
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 return ret;
1439}
1440#endif
1441
1442#ifdef CONFIG_HOTPLUG_CPU
1443/* It's optimal to keep kswapds on the same CPUs as their memory, but
1444 not required for correctness. So if the last cpu in a node goes
1445 away, we get changed to run anywhere: as the first one comes back,
1446 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07001447static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08001448 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449{
1450 pg_data_t *pgdat;
1451 cpumask_t mask;
1452
1453 if (action == CPU_ONLINE) {
KAMEZAWA Hiroyukiec936fc2006-03-27 01:15:59 -08001454 for_each_online_pgdat(pgdat) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455 mask = node_to_cpumask(pgdat->node_id);
1456 if (any_online_cpu(mask) != NR_CPUS)
1457 /* One of our CPUs online: restore mask */
1458 set_cpus_allowed(pgdat->kswapd, mask);
1459 }
1460 }
1461 return NOTIFY_OK;
1462}
1463#endif /* CONFIG_HOTPLUG_CPU */
1464
Yasunori Goto3218ae12006-06-27 02:53:33 -07001465/*
1466 * This kswapd start function will be called by init and node-hot-add.
1467 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1468 */
1469int kswapd_run(int nid)
1470{
1471 pg_data_t *pgdat = NODE_DATA(nid);
1472 int ret = 0;
1473
1474 if (pgdat->kswapd)
1475 return 0;
1476
1477 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1478 if (IS_ERR(pgdat->kswapd)) {
1479 /* failure at boot is fatal */
1480 BUG_ON(system_state == SYSTEM_BOOTING);
1481 printk("Failed to start kswapd on node %d\n",nid);
1482 ret = -1;
1483 }
1484 return ret;
1485}
1486
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487static int __init kswapd_init(void)
1488{
Yasunori Goto3218ae12006-06-27 02:53:33 -07001489 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08001490
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491 swap_setup();
Yasunori Goto3218ae12006-06-27 02:53:33 -07001492 for_each_online_node(nid)
1493 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 hotcpu_notifier(cpu_callback, 0);
1495 return 0;
1496}
1497
1498module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08001499
1500#ifdef CONFIG_NUMA
1501/*
1502 * Zone reclaim mode
1503 *
1504 * If non-zero call zone_reclaim when the number of free pages falls below
1505 * the watermarks.
1506 *
1507 * In the future we may add flags to the mode. However, the page allocator
1508 * should only have to check that zone_reclaim_mode != 0 before calling
1509 * zone_reclaim().
1510 */
1511int zone_reclaim_mode __read_mostly;
1512
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08001513#define RECLAIM_OFF 0
1514#define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1515#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1516#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001517#define RECLAIM_SLAB (1<<3) /* Do a global slab shrink if the zone is out of memory */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08001518
Christoph Lameter9eeff232006-01-18 17:42:31 -08001519/*
Christoph Lametera92f7122006-02-01 03:05:32 -08001520 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1521 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1522 * a zone.
1523 */
1524#define ZONE_RECLAIM_PRIORITY 4
1525
Christoph Lameter9eeff232006-01-18 17:42:31 -08001526/*
1527 * Try to free up some pages from this zone through reclaim.
1528 */
Andrew Morton179e9632006-03-22 00:08:18 -08001529static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08001530{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08001531 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08001532 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001533 struct task_struct *p = current;
1534 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08001535 int priority;
Andrew Morton05ff5132006-03-22 00:08:20 -08001536 unsigned long nr_reclaimed = 0;
Andrew Morton179e9632006-03-22 00:08:18 -08001537 struct scan_control sc = {
1538 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1539 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
Andrew Morton69e05942006-03-22 00:08:19 -08001540 .swap_cluster_max = max_t(unsigned long, nr_pages,
1541 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08001542 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001543 .swappiness = vm_swappiness,
Andrew Morton179e9632006-03-22 00:08:18 -08001544 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08001545
1546 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08001547 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08001548 /*
1549 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1550 * and we also need to be able to write out pages for RECLAIM_WRITE
1551 * and RECLAIM_SWAP.
1552 */
1553 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001554 reclaim_state.reclaimed_slab = 0;
1555 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08001556
Christoph Lametera92f7122006-02-01 03:05:32 -08001557 /*
1558 * Free memory by calling shrink zone with increasing priorities
1559 * until we have enough memory freed.
1560 */
Christoph Lameter86959492006-03-22 00:08:18 -08001561 priority = ZONE_RECLAIM_PRIORITY;
Christoph Lametera92f7122006-02-01 03:05:32 -08001562 do {
Andrew Morton05ff5132006-03-22 00:08:20 -08001563 nr_reclaimed += shrink_zone(priority, zone, &sc);
Christoph Lameter86959492006-03-22 00:08:18 -08001564 priority--;
Andrew Morton05ff5132006-03-22 00:08:20 -08001565 } while (priority >= 0 && nr_reclaimed < nr_pages);
Christoph Lameterc84db232006-02-01 03:05:29 -08001566
Andrew Morton05ff5132006-03-22 00:08:20 -08001567 if (nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001568 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08001569 * shrink_slab() does not currently allow us to determine how
1570 * many pages were freed in this zone. So we just shake the slab
1571 * a bit and then go off node for this particular allocation
1572 * despite possibly having freed enough memory to allocate in
1573 * this zone. If we freed local memory then the next
1574 * allocations will be local again.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001575 *
1576 * shrink_slab will free memory on all zones and may take
1577 * a long time.
1578 */
1579 shrink_slab(sc.nr_scanned, gfp_mask, order);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08001580 }
1581
Christoph Lameter9eeff232006-01-18 17:42:31 -08001582 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08001583 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Andrew Morton05ff5132006-03-22 00:08:20 -08001584 return nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001585}
Andrew Morton179e9632006-03-22 00:08:18 -08001586
1587int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1588{
1589 cpumask_t mask;
1590 int node_id;
1591
1592 /*
Christoph Lameter34aa1332006-06-30 01:55:37 -07001593 * Do not reclaim if there are not enough reclaimable pages in this
1594 * zone that would satify this allocations.
1595 *
1596 * All unmapped pagecache pages are reclaimable.
1597 *
1598 * Both counters may be temporarily off a bit so we use
1599 * SWAP_CLUSTER_MAX as the boundary. It may also be good to
1600 * leave a few frequently used unmapped pagecache pages around.
Andrew Morton179e9632006-03-22 00:08:18 -08001601 */
Christoph Lameter34aa1332006-06-30 01:55:37 -07001602 if (zone_page_state(zone, NR_FILE_PAGES) -
1603 zone_page_state(zone, NR_FILE_MAPPED) < SWAP_CLUSTER_MAX)
Andrew Morton179e9632006-03-22 00:08:18 -08001604 return 0;
1605
1606 /*
1607 * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1608 * not have reclaimable pages and if we should not delay the allocation
1609 * then do not scan.
1610 */
1611 if (!(gfp_mask & __GFP_WAIT) ||
1612 zone->all_unreclaimable ||
1613 atomic_read(&zone->reclaim_in_progress) > 0 ||
1614 (current->flags & PF_MEMALLOC))
1615 return 0;
1616
1617 /*
1618 * Only run zone reclaim on the local zone or on zones that do not
1619 * have associated processors. This will favor the local processor
1620 * over remote processors and spread off node memory allocations
1621 * as wide as possible.
1622 */
1623 node_id = zone->zone_pgdat->node_id;
1624 mask = node_to_cpumask(node_id);
1625 if (!cpus_empty(mask) && node_id != numa_node_id())
1626 return 0;
1627 return __zone_reclaim(zone, gfp_mask, order);
1628}
Christoph Lameter9eeff232006-01-18 17:42:31 -08001629#endif